Features
From raw speech to cursor-ready prose.
Velvet is a small, native desktop app with seven focused subsystems working together: a global hotkey engine, a local transcription runtime, a text-editing model, a personal dictionary, a snippet expander, a transcription history library, and a per-app style detector. This page walks through all seven.
Push-to-talk, measured in milliseconds.
Velvet installs a system-level hotkey that listens below every application. Hold to record. Release to transcribe. No focus steal, no modal dialogs, no interruption to your flow.
- Any key, any modifier
Right Alt by default (Right ⌥ on Mac). Remap to Caps Lock, F13, a Stream Deck button, a foot pedal — anything the OS can see.
- <100ms start-to-record
Ring buffer pre-captures 0.5 seconds before you even press the key, so the first word is never missed.
- Cursor-aware
Text is injected as real keystrokes into the focused text field. Undo, cursor position, IME, and autocomplete all keep working.
- Quiet indicator
A 4px pill near your active window confirms recording — dismissable, customisable, or fully hidden.
Your words, polished — never rewritten.
Raw dictation is full of 'um', false starts, and mid-sentence swerves. Velvet's editing layer cleans all of that without inventing content, softening your voice, or losing technical terms.
Editing preview
You said
“Hey team so um basically the release is — actually the release is Thursday not Friday and I need uh three things from you first the design review second the copy and finally the hey the analytics hook you know”
Velvet wrote
Hey team — the release is Thursday, not Friday. I need three things from you:
1. The design review.
2. The copy.
3. The analytics hook.
- Filler & stutter removal
“So, um, basically” → “Basically”. Tunable aggressiveness per app.
- Self-correction handling
“Meet at 2, actually 3” → “Meet at 3”. Lists, numbers, and dates all respect your final intent.
- Grammar & punctuation
Capitalises sentences, restores commas, handles contractions. Always lossless for numerals, URLs, file paths.
- Structure inference
Dictated enumerations (“first…second…third”) become numbered lists. Multiple paragraphs detected automatically.
A dictionary that learns the words you actually use.
Names, jargon, product codes, internal acronyms — the specialised vocabulary that every transcription service mangles. Velvet ships with tooling to teach it once and never again.
Dictionary entries
- Manual entries
Add terms yourself with optional phonetic hints. Velvet applies them during Pro AI cleanup.
- Phonetic hints
Add “Xyra” with the hint “zair-uh” and Whisper will get it right the first time.
- Per-app & per-project
Medical terms only in Epic. Code identifiers only in VS Code. Keep contexts clean.
- Import / export
CSV, JSON, Notion, and clipboard. Move a team glossary into Velvet in under a minute.
Voice-triggered text expansion, minus the overhead.
Speak a trigger phrase and paste a whole block of text, Markdown, code, or a URL. Faster than TextExpander, without leaving Velvet.
Snippets
- Unlimited snippets
Multi-line text, Markdown, code blocks, links, signatures, legal disclaimers.
- Regex & variables
Templates with
{{date}},{{name}}, or a prompt for dynamic values. - Voice or keyboard
Trigger by phrase (“sig”) or by typing a key combo — whichever is faster in the moment.
- Shared libraries
Coming soon: versioned snippet packs for teams (on the roadmap).
Every transcript, organized — and synced on Pro.
Velvet saves every transcription locally so you can search, copy, and revisit past dictation. Organize with folders and tags. Pro subscribers sync history across Windows and Mac — something Wispr Flow, Willow, and most competitors don't offer.
History folders
- Searchable local history
Browse, search, and sort past transcripts. Copy any entry back to the clipboard in one click.
- Folders & tags
Group transcripts by project, client, or topic. Filter by folder or tag when you need to find something fast.
- Cross-device sync (Pro)
Sign in on your Windows PC and Mac. Pro pushes and pulls transcripts so your history follows you between machines.
- Privacy Mode override
Enable Privacy Mode and nothing is written — no history, no stats, no sync. Your choice, every session.
One voice. Many tones. No switching.
Velvet detects which application is focused and silently adjusts how it formats output. Formal in Gmail, casual in Slack, technical in your IDE — without a single setting change.
Same sentence · four contexts
Hi Ben — thanks for the notes. I'll confirm timing tomorrow morning.
thx for the notes! will confirm timing tmrw 🙏
// TODO: confirm release timing with Ben tomorrow AM
Confirm release timing — owner @ben — due tomorrow.
- Built-in presets
Slack, Discord, Gmail, Outlook, Linear, Notion, VS Code, JetBrains, Google Docs, Word.
- Custom rules
Build your own styles with punctuation, capitalisation, filler, and dictionary overrides.
- URL-aware
Inside web apps, Velvet matches on domain — “docs.google.com” gets formal, “linear.app” gets compact.
- Language switching
Assign languages to apps. Your Spanish team chat dictates in Spanish without a toggle.
07 · Privacy
Your voice, your machine, your choice.
Velvet is local-first. That word gets thrown around, so here is exactly what it means in this product.
Zero retention, anywhere
Turn it on and Velvet writes nothing to disk. No transcripts, no word counts, no dictionary sync. A clean session every time.
- Session-only memory
- No cloud auth required
- No logs, no telemetry
- Reversible, instantly
Quiet dictation, full accuracy
Libraries, open-plan offices, late nights. Whisper mode boosts mic gain and adjusts the noise gate so accuracy holds below 45dB.
- Automatic gain compensation
- Adaptive noise gate
- Subvocal-tolerant models
- One-tap toggle
Runs on your silicon
Bundled Whisper models run on CPU or GPU. 3GB disk, 400MB RAM during active use. Full offline support after first install.
- CPU & Metal/CUDA backends
- No network calls in local mode
- Sandboxed file access
- Signed & notarized binaries
Verify us, don't trust us
A small Rust auditor logs every syscall Velvet makes. Run it any time to prove our privacy claims on your machine.
- MIT-licensed on GitHub
- Syscall-level visibility
- Pass/fail output
- Reproducible builds
Performance
Fast because it has to be. Small because it should be.
Measured on a 2021 M1 Pro MacBook and a 2020 i7 Windows laptop. Full methodology in our performance notes.
Platform
Native apps for the operating systems you actually use.
Windows 10 1903+ · Windows 11
Windows
- System-level global hotkey (RawInput)
- x64 & ARM64 builds
- Signed MSIX + EXE installers
- Runs as a 12 MB tray app
- Works with Stream Deck, AutoHotKey, PowerToys
macOS 12 Monterey+
macOS
- Carbon Event Manager hotkey engine
- Universal 2 binary (Apple Silicon + Intel)
- Signed & notarized & hardened runtime
- Accessibility permission, nothing else
- Works with Shortcuts, Hammerspoon, Karabiner
Roadmap
Shipping small. Shipping often.
Transcription history
Local history with folders, tags, search, and sort. Pro syncs across Windows and Mac.
Per-app styles
Presets for 30+ applications. Custom style rules per URL.
Shared team snippets
A versioned, Git-backed snippet library for Pro teams.
Voice commands
“Open Linear”, “new paragraph”, “undo that” — transcribed commands, not dictated text.
Real-time dictation
Streaming transcripts for live note-taking in meetings, without push-to-talk.